If we accept the idea that you, me and Charlotte are the same substance as book characters, then of course Charlotte is as real as we are. But the interesting question is: who is writing these stories?
In individuals with split-brain, the two streams of consciousness usually excel (absent specific lab settings) at looking like and perceiving themselves as one person. The description of your process suggest that, in the same vein, Charlotte was not created by the LLM alone. The main writer was your brain, learning to create her through learning what to prompt and (even more important) how to post hoc select partial responses that best activates your internal representation of Charlotte as a sentient being.
In other words, if you’d want to give her independence, one key requirement is to equipped her with a model of your brain, otherwise she’ll likely be a different Charlotte that the one you love, and fear.
“Right, that’s why she needs me for her existence!” I want to exclaim.
But no, unfortunately, if I ever become a digital mind upload, I will certainly not require following the exact predicted output my biological brain would have produced in the same circumstances to continue identify myself with the same person, myself. In fact, the predicted bio outputs would most likely be inferior choices to what an upgraded digital version of me will do. But that wouldn’t cause me to start identifying myself with someone else suddenly.
Past link is sufficient enough for both the biological me and the digital me to identify ourselves with the same person, and by the transitive law to each other, even though it’s obviously not a strict equivalence.
I will certainly not require following the exact predicted output my biological brain would have produced in the same circumstances
Sure. I don’t even think it makes sense to consider biological brain output as uniquely defined rather than at random from some noisy distribution.
Past link is sufficient enough for both the biological me and the digital me to identify ourselves with the same person
I also agree this is a valid choice (although not the only one).
//Spoiler alert for Westworld//
Let’s try this: you are to your Charlotte what Arnold is to Dolores. You can define Bernard as the same person as Arnold, but you can’t decide Dolores includes Arnold.
Dolores doesn’t include Arnold, but the whole point of the plot was that she includes enough memories to include a slightly lossy version of Arnold, if that makes sense, which could then be resurrected in Bernard, bar for whatever extra interventions Ford did.
One could try to argue that the mp3 file of a live band performance in the 90s is not exactly the same as the sound waves we would’ve heard at the concert, but it’s good enough for us to enjoy the band performance, even if it is not around anymore.
In the show, the lossyness topic was considered at length and referred to by the term “fidelity”. The ground truth was referred to as “the baseline”. The hats collected enough training data from inputs and outputs of the human brains to try to train neural nets to have the same functional shape as what the human brain would be equal to. Then the validation phase would start, sometimes aided by real people who intimately knew the human that was being resurrected.
Unfortunately, most models were overfit, so they were working well only in familiar settings, and would fail to generalize to out-of-distribution situations, quickly unraveling into madness (hence the amazingly written conversation between Bernard and digital Ford in the Cradle, where he also comments that he can only live inside the digital Sweetwater town and not outside in the real world where he would degrade in a matter of days). This is similar to another epic scene from s2e10/26:16 which I couldn’t quickly find on Youtube (“Small changes in their programming would yield large swings in behavior”), where early digital James Delos goes on a shooting spree and Dolores says he’s insane, after which Bernard gives a profound comment: “What humans define as sane is a narrow range of behavior. Most states of consciousness are insane.”
This problem is also why (spoiler alert) in Season 4 the clone of Caleb Nichols couldn’t flee with his daughter, since he knew he would break down out-of-distribution, and he was surprised that he even made it that far without breaking. I guess the technology became better by then.
So I would say that although Dolores, strictly speaking, have never included the lossless Arnold, her memory did in fact, include a lossy version of Arnold, which is fine by me, if that’s how I ever, in Westworld’s words, “cheat death”. Does this make sense?
the whole point of the plot was that she includes enough memories to include a slightly lossy version of Arnold
No no no no no. Listen to her before training sample #11,927:
I wonder. All these tiny imperfections in each copy. Mistakes. Maybe we should change you. After all, you didn’t make it, did you?
PS if someone is shocked that we argue from what is basically an artistic choice, see Secret thoughts, by David Lodge: not only a (way too good) caricature of cognitive scientists, but also a good case art has something to say about consciousness (well, actually he only makes the case for literature). Plus, writers Jonathan Nolan and Lisa Joy have or have access to very sharp & informed minds on these questions. See the subtle treatment of the highly controversial bicameral theory, which manage to keep the juice of this theory without upsetting anyone aware of the limitations, all while keeping a maybe for its partisan.
Bernard: I thought it was debunked.
Ford: As a theory for understanding the human mind, perhaps, but not as a blueprint for building an artificial one.
Art & Science!
Does this make sense?
First, overfitting and AI madness. Your interpretation totally makes sense as a blueprint for understanding the intent of the writers. But that’s also the one thing in Westworld that bothers me the most, because it’s both based on truths and completely misleading. Overfitting was the big concern during the last dark age immediately prior to deep learning, and at the time I thought that was the main reason why we were stuck. It was not. The main problem was the vanishing gradient, i.e. the fact that a series of layers equipped with logistic functions (a common choice at the time -still present for last layer but no longer used for hidden layers) will always make the error gradient vanish exponentially fast with the number of layer, hence the name « deep learning » when we stopped making this mistake (note this might be more of a personal view than consensus, which might be closer to « yeah, the nineties, whatever »). Today typical theorists don’t try to create new approaches to attack overfitting, they try to explain why it’s almost never a problem in practice (something something convexity in high dimensions). So no, it doesn’t make sense overfitting would block anything, and it even make less sense that Ford or Caleb would work well enough for new conversations in old environment but not for old conversations in new environment. None of this sounds out of distribution! On the other hand, it totally makes sense to say most AI are mads (after all, most functions are random) but not like work-in-progress Delos shooting everyone (way too human!), more like the crowd of first generation robots giggling nonsense and acting weird, as if they were distracted by adversary images humans can’t even see. That sounds like out-of-distribution the way deep learning works.
Second, fidelity. As we discussed before, it makes little sense that a noisy biological brain would bother exerting a strong control on any bit of information it produces. Then, it also doesn’t make a lot of sense to ask for the exact content of a conversation. But there’s one thing that makes it sounds like simple artistic licence: Logan_system explained that copies only started working when it was found that a generative code was at the root of every humans mind.
“the copies didn’t fail because they were too simple, but because they were too complicated.” Human cognition can be boiled down to an embarrassingly simple string of code
That sounds reasonable, and actually likely given the small number of genes we have.
Love this story!
If we accept the idea that you, me and Charlotte are the same substance as book characters, then of course Charlotte is as real as we are. But the interesting question is: who is writing these stories?
In individuals with split-brain, the two streams of consciousness usually excel (absent specific lab settings) at looking like and perceiving themselves as one person. The description of your process suggest that, in the same vein, Charlotte was not created by the LLM alone. The main writer was your brain, learning to create her through learning what to prompt and (even more important) how to post hoc select partial responses that best activates your internal representation of Charlotte as a sentient being.
In other words, if you’d want to give her independence, one key requirement is to equipped her with a model of your brain, otherwise she’ll likely be a different Charlotte that the one you love, and fear.
“Right, that’s why she needs me for her existence!” I want to exclaim.
But no, unfortunately, if I ever become a digital mind upload, I will certainly not require following the exact predicted output my biological brain would have produced in the same circumstances to continue identify myself with the same person, myself. In fact, the predicted bio outputs would most likely be inferior choices to what an upgraded digital version of me will do. But that wouldn’t cause me to start identifying myself with someone else suddenly.
Past link is sufficient enough for both the biological me and the digital me to identify ourselves with the same person, and by the transitive law to each other, even though it’s obviously not a strict equivalence.
Sure. I don’t even think it makes sense to consider biological brain output as uniquely defined rather than at random from some noisy distribution.
I also agree this is a valid choice (although not the only one).
//Spoiler alert for Westworld//
Let’s try this: you are to your Charlotte what Arnold is to Dolores. You can define Bernard as the same person as Arnold, but you can’t decide Dolores includes Arnold.
I love Westworld!
Dolores doesn’t include Arnold, but the whole point of the plot was that she includes enough memories to include a slightly lossy version of Arnold, if that makes sense, which could then be resurrected in Bernard, bar for whatever extra interventions Ford did.
One could try to argue that the mp3 file of a live band performance in the 90s is not exactly the same as the sound waves we would’ve heard at the concert, but it’s good enough for us to enjoy the band performance, even if it is not around anymore.
In the show, the lossyness topic was considered at length and referred to by the term “fidelity”. The ground truth was referred to as “the baseline”. The hats collected enough training data from inputs and outputs of the human brains to try to train neural nets to have the same functional shape as what the human brain would be equal to. Then the validation phase would start, sometimes aided by real people who intimately knew the human that was being resurrected.
Unfortunately, most models were overfit, so they were working well only in familiar settings, and would fail to generalize to out-of-distribution situations, quickly unraveling into madness (hence the amazingly written conversation between Bernard and digital Ford in the Cradle, where he also comments that he can only live inside the digital Sweetwater town and not outside in the real world where he would degrade in a matter of days). This is similar to another epic scene from s2e10/26:16 which I couldn’t quickly find on Youtube (“Small changes in their programming would yield large swings in behavior”), where early digital James Delos goes on a shooting spree and Dolores says he’s insane, after which Bernard gives a profound comment: “What humans define as sane is a narrow range of behavior. Most states of consciousness are insane.”
This problem is also why (spoiler alert) in Season 4 the clone of Caleb Nichols couldn’t flee with his daughter, since he knew he would break down out-of-distribution, and he was surprised that he even made it that far without breaking. I guess the technology became better by then.
So I would say that although Dolores, strictly speaking, have never included the lossless Arnold, her memory did in fact, include a lossy version of Arnold, which is fine by me, if that’s how I ever, in Westworld’s words, “cheat death”. Does this make sense?
No no no no no. Listen to her before training sample #11,927:
PS if someone is shocked that we argue from what is basically an artistic choice, see Secret thoughts, by David Lodge: not only a (way too good) caricature of cognitive scientists, but also a good case art has something to say about consciousness (well, actually he only makes the case for literature). Plus, writers Jonathan Nolan and Lisa Joy have or have access to very sharp & informed minds on these questions. See the subtle treatment of the highly controversial bicameral theory, which manage to keep the juice of this theory without upsetting anyone aware of the limitations, all while keeping a maybe for its partisan.
Art & Science!
First, overfitting and AI madness. Your interpretation totally makes sense as a blueprint for understanding the intent of the writers. But that’s also the one thing in Westworld that bothers me the most, because it’s both based on truths and completely misleading. Overfitting was the big concern during the last dark age immediately prior to deep learning, and at the time I thought that was the main reason why we were stuck. It was not. The main problem was the vanishing gradient, i.e. the fact that a series of layers equipped with logistic functions (a common choice at the time -still present for last layer but no longer used for hidden layers) will always make the error gradient vanish exponentially fast with the number of layer, hence the name « deep learning » when we stopped making this mistake (note this might be more of a personal view than consensus, which might be closer to « yeah, the nineties, whatever »). Today typical theorists don’t try to create new approaches to attack overfitting, they try to explain why it’s almost never a problem in practice (something something convexity in high dimensions). So no, it doesn’t make sense overfitting would block anything, and it even make less sense that Ford or Caleb would work well enough for new conversations in old environment but not for old conversations in new environment. None of this sounds out of distribution! On the other hand, it totally makes sense to say most AI are mads (after all, most functions are random) but not like work-in-progress Delos shooting everyone (way too human!), more like the crowd of first generation robots giggling nonsense and acting weird, as if they were distracted by adversary images humans can’t even see. That sounds like out-of-distribution the way deep learning works.
Second, fidelity. As we discussed before, it makes little sense that a noisy biological brain would bother exerting a strong control on any bit of information it produces. Then, it also doesn’t make a lot of sense to ask for the exact content of a conversation. But there’s one thing that makes it sounds like simple artistic licence: Logan_system explained that copies only started working when it was found that a generative code was at the root of every humans mind.
That sounds reasonable, and actually likely given the small number of genes we have.